add gridworld (and a parent class 'BinaryInputSampler')
Browse files- automata.py +58 -11
automata.py
CHANGED
@@ -21,6 +21,13 @@ import itertools
|
|
21 |
|
22 |
import datasets
|
23 |
import numpy as np
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
# Local imports
|
26 |
# from symmetric import SymmetricSampler
|
@@ -52,9 +59,9 @@ class SyntheticAutomataDataset(datasets.GeneratorBasedBuilder):
|
|
52 |
"""
|
53 |
if 'name' not in config:
|
54 |
config['name'] = 'parity'
|
55 |
-
if 'length' not in config:
|
56 |
config['length'] = 20
|
57 |
-
if 'size' not in config:
|
58 |
config['size'] = -1
|
59 |
|
60 |
self.data_config = config
|
@@ -119,25 +126,65 @@ class AutomatonSampler:
|
|
119 |
raise NotImplementedError()
|
120 |
|
121 |
|
122 |
-
class
|
123 |
def __init__(self, data_config):
|
124 |
super().__init__(data_config)
|
125 |
-
|
126 |
-
|
|
|
|
|
127 |
|
128 |
def f(self, x):
|
129 |
-
|
130 |
|
131 |
def sample(self):
|
132 |
-
x = self.np_rng.binomial(1,
|
133 |
return x, self.f(x)
|
134 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
|
136 |
class FlipFlopSampler(AutomatonSampler):
|
137 |
def __init__(self, data_config):
|
138 |
super().__init__(data_config)
|
139 |
self.name = 'flipflop'
|
140 |
-
self.data_config = data_config
|
141 |
|
142 |
if 'n' not in data_config:
|
143 |
data_config['n'] = 2
|
@@ -170,7 +217,6 @@ class SymmetricSampler(AutomatonSampler):
|
|
170 |
def __init__(self, data_config):
|
171 |
super().__init__(data_config)
|
172 |
self.name = 'symmetric'
|
173 |
-
self.data_config = data_config
|
174 |
|
175 |
if 'n' not in data_config:
|
176 |
data_config['n'] = 5 # Default to S5
|
@@ -180,7 +226,7 @@ class SymmetricSampler(AutomatonSampler):
|
|
180 |
# Options: 'state', 'first_chair'
|
181 |
data_config['label_type'] = 'state'
|
182 |
|
183 |
-
self.n = data_config['n']
|
184 |
self.label_type = data_config['label_type']
|
185 |
|
186 |
"""
|
@@ -242,8 +288,9 @@ class SymmetricSampler(AutomatonSampler):
|
|
242 |
|
243 |
|
244 |
dataset_map = {
|
245 |
-
'
|
246 |
'flipflop': FlipFlopSampler,
|
|
|
247 |
'symmetric': SymmetricSampler,
|
248 |
# TODO: more datasets
|
249 |
}
|
|
|
21 |
|
22 |
import datasets
|
23 |
import numpy as np
|
24 |
+
from copy import copy
|
25 |
+
|
26 |
+
# check python version
|
27 |
+
import sys
|
28 |
+
major, minor = sys.version_info[:2]
|
29 |
+
version = major + 0.1*minor
|
30 |
+
OLD_PY_VERSION = 1 if version < 3.8 else 0
|
31 |
|
32 |
# Local imports
|
33 |
# from symmetric import SymmetricSampler
|
|
|
59 |
"""
|
60 |
if 'name' not in config:
|
61 |
config['name'] = 'parity'
|
62 |
+
if 'length' not in config: # sequence length
|
63 |
config['length'] = 20
|
64 |
+
if 'size' not in config: # number of sequences
|
65 |
config['size'] = -1
|
66 |
|
67 |
self.data_config = config
|
|
|
126 |
raise NotImplementedError()
|
127 |
|
128 |
|
129 |
+
class BinaryInputSampler(AutomatonSampler):
|
130 |
def __init__(self, data_config):
|
131 |
super().__init__(data_config)
|
132 |
+
|
133 |
+
if 'prob1' not in data_config:
|
134 |
+
data_config['prob1'] = 0.5
|
135 |
+
self.prob1 = data_config['prob1']
|
136 |
|
137 |
def f(self, x):
|
138 |
+
raise NotImplementedError()
|
139 |
|
140 |
def sample(self):
|
141 |
+
x = self.np_rng.binomial(1, self.prob1, size=self.T)
|
142 |
return x, self.f(x)
|
143 |
|
144 |
+
class ParitySampler(BinaryInputSampler):
|
145 |
+
def __init__(self, data_config):
|
146 |
+
super().__init__(data_config)
|
147 |
+
self.name = 'parity'
|
148 |
+
|
149 |
+
def f(self, x):
|
150 |
+
return np.cumsum(x) % 2
|
151 |
+
|
152 |
+
class GridworldSampler(BinaryInputSampler):
|
153 |
+
"""
|
154 |
+
Note: gridworld currently doesn't include a no-op.
|
155 |
+
"""
|
156 |
+
def __init__(self, data_config):
|
157 |
+
super().__init__(data_config)
|
158 |
+
self.name = 'gridworld'
|
159 |
+
|
160 |
+
if 'n' not in data_config:
|
161 |
+
data_config['n'] = 9
|
162 |
+
"""
|
163 |
+
NOTE: n is the number of states, and S is the id (0-indexing) of the rightmost state.
|
164 |
+
i.e. the states are 0,1,2,...,S, where S=n-1.
|
165 |
+
"""
|
166 |
+
self.n = data_config['n']
|
167 |
+
self.S = self.n - 1
|
168 |
+
|
169 |
+
def f(self, x):
|
170 |
+
x = copy(x)
|
171 |
+
x[x == 0] = -1
|
172 |
+
if OLD_PY_VERSION:
|
173 |
+
# NOTE: for Python 3.7 or below, accumulate doesn't have the 'initial' argument.
|
174 |
+
x = np.concatenate([np.array([0]), x]).astype(np.int64)
|
175 |
+
states = list(itertools.accumulate(x, lambda a,b: max(min(a+b, self.S), 0)))
|
176 |
+
states = states[1:]
|
177 |
+
else:
|
178 |
+
states = list(itertools.accumulate(x, lambda a,b: max(min(a+b, self.S), 0), initial=0))
|
179 |
+
states = states[1:] # remove the 1st entry with is the (meaningless) initial value 0
|
180 |
+
return np.array(states).astype(np.int64)
|
181 |
+
|
182 |
+
|
183 |
|
184 |
class FlipFlopSampler(AutomatonSampler):
|
185 |
def __init__(self, data_config):
|
186 |
super().__init__(data_config)
|
187 |
self.name = 'flipflop'
|
|
|
188 |
|
189 |
if 'n' not in data_config:
|
190 |
data_config['n'] = 2
|
|
|
217 |
def __init__(self, data_config):
|
218 |
super().__init__(data_config)
|
219 |
self.name = 'symmetric'
|
|
|
220 |
|
221 |
if 'n' not in data_config:
|
222 |
data_config['n'] = 5 # Default to S5
|
|
|
226 |
# Options: 'state', 'first_chair'
|
227 |
data_config['label_type'] = 'state'
|
228 |
|
229 |
+
self.n = data_config['n'] # the symmetric group Sn
|
230 |
self.label_type = data_config['label_type']
|
231 |
|
232 |
"""
|
|
|
288 |
|
289 |
|
290 |
dataset_map = {
|
291 |
+
'gridworld': GridworldSampler,
|
292 |
'flipflop': FlipFlopSampler,
|
293 |
+
'parity': ParitySampler,
|
294 |
'symmetric': SymmetricSampler,
|
295 |
# TODO: more datasets
|
296 |
}
|